Debiasing matrix completion via first estimating missingness probabilities via matrix completion

This code accompanies our NeurIPS 2019 paper "Missing Not at Random in Matrix Completion: The Effectiveness of Estimating Missingness Probabilities under a Low Nuclear Norm Assumption".

Authors: George H. Chen (georgechen@cmu.edu), Wei Ma (wei.w.ma@polyu.edu.hk)

We have also included code by some other authors, namely:

• ExpoMF (Liang et al 2016): https://github.com/dawenl/expo-mf
• A Python implemention by github user andrewdalex of SoftImpute-ALS (Hastie et al 2015): https://github.com/andrewdalex/SoftImpute-ALS

Code dependencies

We tested this code using Anaconda Python 3.7 in a Linux environment (Ubuntu 18.04) with these additional packages:

• surprise (install using pip install -U surprise)
• copt (install using pip install -U copt)
• hnswlib (install using pip install hnswlib)

We modified Surprise's SVD and SVDpp to allow for weighted entries, and we also have helper functions coded in cython; these require cython compilation:

python setup_mnar_mc_helpers.py build_ext --inplace

Getting the code to run

To run the code, you must first prepare datasets and then you can run python demo.py <dataset name> (edit demo.py to specify which matrix completion algorithms to run).

For the synthetic datasets, prepare them by running python prepare_synthetic.py.

Then you should be able to run python demo.py steck-0 (as well as steck-1, steck-2, up through steck-9 for the MovieLoverData and useritemfeature-0, useritemfeature-1, up through useritemfeature-9 for the UserItemData).

For the Coat dataset, download it here: https://www.cs.cornell.edu/~schnabts/mnar/

Copy it to ./coat/

Run python prepare_coat.py. Then you should be able to run python demo.py coat.

For the MovieLens-100k dataset, download it here: https://grouplens.org/datasets/movielens/100k/

Copy it to ./ml-100k/

Run python prepare_ml100k.py ml-100k-0 (as well as ml-100k-1, ml-100k-2, up through ml-100k-9)